Device-SMI

Self-contained Python lib with zero-dependencies that give you a unified `device` properties for `gpu`, `cpu`, and `npu`. No more calling separate tools such as `nvidia-smi` or `/proc/cpuinfo` and parsing it yourself.

News

• 10/01/2025 0.5.0 Fast (thread interval polled) .metrics(fast=True) for low-latency close-to-real-time device resource usage.
• 03/01/2025 0.4.1 Fix compat with AMD ROCm 6.3.2 and MI300X.
• 02/26/2025 0.4.0 Added AMD GPU support. Validated with amd-smi on 7900XTX. Use PowerShell for auto-device selection on Windows platform.
• 12/20/2024 0.3.3 Patch fix for Windows install compatibility.
• 12/05/2024 0.3.2 Added Windows WSL support.
• 12/03/2024 0.3.1 Added CPUDevice compat for Windows.
• 12/02/2024 0.3.0 Added OSDevice.[name, version, kernel, arch] for Linux/MacOS/Windows/FreeBSD/Solaris. Added cpu.[count, cores, threads] properties. Bug fix for gpu device index.
• 11/29/2024 0.2.1 Added pcie.[gen, speed, id] + [firmware, driver] properties to GPU device.

Features

• Retrieve information for both CPU and GPU devices.
• Includes details about memory usage, utilization, driver, pcie info when applicable, and other device specifications.
• Zero pypi dependency.
• Linux/MacOS support

Supported Devices:

• OS: Linux, MacOS, Windows, FreeBSD, Solaris
• CPU: [Intel/AMD/Apple] Linux/MacOS system interface
• NVIDIA GPU: NVIDIA System Management Interface nvidia-smi
• Intel XPU: Intel/XPU System Management Interface xpu-smi
• AMD ROCm/GPU: AMD System Management Interface amd-smi
• Apple GPU: MacOS interfaces

Usage

For OS, use os to init a new Device object.

from device_smi import Device

dev = Device("os")
print(dev)

Output: (Ubuntu 22.04)

{'type': 'os', 'name': 'ubuntu', 'version': '22.04', 'kernel': '6.12.1-x64v3-xanmod2', 'arch': 'x86_64'}

For GPU/XPU, use [gpu, cuda] for Nvidia and xpu for Intel/XPU. Index usage for multiple GPUs: cuda:0

from device_smi import Device

dev = Device("cuda:0")
print(dev)

Output: (A100)

{'memory_total': 103079215104, 'type': 'gpu', 'features': ['8.0'], 'vendor': 'nvidia', 'model': 'pg506-230', 'pcie': {'gen': 4, 'speed': 4, 'id': '00000000:21:00.0'}, 'gpu': {'driver': '580.82.07', 'firmware': '92.00.4F.00.01'}}

For CPU, use cpu to init a new Device object.

from device_smi import Device

dev = Device("cpu")
print(dev)

Output: (AMD EPYC 7443)

{'memory_total': 2151497490432, 'type': 'cpu', 'features': ['3dnowprefetch', 'abm', 'adx', 'aes', 'amd_ppin', 'aperfmperf', 'apic', 'arat', 'avx', 'avx2', 'bmi1', 'bmi2', 'bpext', 'brs', 'cat_l3', 'cdp_l3', 'clflush', 'clflushopt', 'clwb', 'clzero', 'cmov', 'cmp_legacy', 'constant_tsc', 'cpb', 'cpuid', 'cqm', 'cqm_llc', 'cqm_mbm_local', 'cqm_mbm_total', 'cqm_occup_llc', 'cr8_legacy', 'cx16', 'cx8', 'de', 'debug_swap', 'decodeassists', 'erms', 'extapic', 'extd_apicid', 'f16c', 'flushbyasid', 'fma', 'fpu', 'fsgsbase', 'fsrm', 'fxsr', 'fxsr_opt', 'ht', 'hw_pstate', 'ibpb', 'ibrs', 'ibs', 'invpcid', 'irperf', 'lahf_lm', 'lbrv', 'lm', 'mba', 'mca', 'mce', 'misalignsse', 'mmx', 'mmxext', 'monitor', 'movbe', 'msr', 'mtrr', 'mwaitx', 'nonstop_tsc', 'nopl', 'npt', 'nrip_save', 'nx', 'ospke', 'osvw', 'overflow_recov', 'pae', 'pat', 'pausefilter', 'pcid', 'pclmulqdq', 'pdpe1gb', 'perfctr_core', 'perfctr_llc', 'perfctr_nb', 'pfthreshold', 'pge', 'pku', 'pni', 'popcnt', 'pse', 'pse36', 'rapl', 'rdpid', 'rdpru', 'rdrand', 'rdseed', 'rdt_a', 'rdtscp', 'rep_good', 'sep', 'sev', 'sev_es', 'sha_ni', 'skinit', 'smap', 'smca', 'smep', 'ssbd', 'sse', 'sse2', 'sse4_1', 'sse4_2', 'sse4a', 'ssse3', 'stibp', 'succor', 'svm', 'svm_lock', 'syscall', 'tce', 'topoext', 'tsc', 'tsc_scale', 'umip', 'user_shstk', 'v_spec_ctrl', 'v_vmsave_vmload', 'vaes', 'vgif', 'vmcb_clean', 'vme', 'vmmcall', 'vpclmulqdq', 'wbnoinvd', 'wdt', 'xgetbv1', 'xsave', 'xsavec', 'xsaveerptr', 'xsaveopt', 'xsaves', 'xtopology'], 'vendor': 'amd', 'model': 'epyc 7443', 'count': 2, 'cores': 48, 'threads': 96}

Runtime metrics

Use device.metric() to fetch the latest utilization snapshot for any supported device. The returned object exposes attributes such as memory_used, memory_process, and utilization.

from device_smi import Device

dev = Device("cuda:0")
stats = dev.metric()
print(stats.memory_used, stats.utilization)

Fast metrics cache

GPU metrics that depend on vendor SMI utilities can be slow. You can opt into a fast, cached version of the runtime data by calling Device.metric(fast=True). Cached data is refreshed in the background every 200ms by default:

from device_smi import Device

dev = Device("cuda:0", fast_metrics_interval=0.200)

stats = dev.metric(fast=True)
print(stats.memory_used, stats.utilization)

NVIDIA NVML and nvidia-smi caching

For NVIDIA devices, Device-SMI first tries to call NVML directly via libnvidia-ml so it does not spawn nvidia-smi repeatedly. When NVML is not available, it falls back to nvidia-smi.

nvidia-smi caching is disabled by default (nvidia_smi_cache_ttl = 0). Set the TTL to a positive value to enable a thread-safe LRU cache that returns stale data within the TTL window:

from device_smi import Device

Device.config.nvidia_smi_cache_ttl = 2.0  # seconds; 0 disables caching
Device.config.nvidia_smi_cache_maxsize = 32

Roadmap

• Support Intel/Gaudi
• Support Google/TPU
• Add NPU support (ARM/Intel/AMD)